Comparison of non-parametric methods in genomic evaluation of discrete traits

Abstract The prediction performance of SVM and Bayesian methods in genomic evaluation of discrete traits was compared in different scenarios of heritability level (0.1, 0.3, and 0.5), number of quantitative trait loci, QTL (60, 300 and 600) and reference population size (1000, 2000 and 5000 individuals). To do this, a genome comprised of 3 chromosomes was simulated on which 6000 bi-allelic SNP markers were distributed. The QTL effects were modeled with gamma distribution was used to model. Pearson's correlation between the true and predicted genomic breeding values was used as the measure of the predictive accuracy. Results showed the significant effect of heritability on prediction accuracy in a way that by increasing heritability, the accuracy of prediction increased significantly. However, the accuracy of prediction was not affected by the number of QTLs. In addition, by increasing the size of the reference population from 1000 to 5000 individuals, the accuracy of prediction increased. Among different kernel functions used to construct SVM, the radial function had higher predictive performance and therefore was used to analysis data. In all the scenarios studied, BayesA outperformed SVM, especially where the heritability level was low. Therefore, the BayesA was recommended for genomic selection of discrete traits. In addition, it was recommended that in different scenarios of genomic selection, different methods should be compared and the best method should be applied for genomic selection.